Confronting Microphysical Uncertainty with BOSS

Abstract:

Representing microphysics within numerical weather and climate models is challenging because we lack fundamental understanding of microphysical processes and are limited by the computational inability to track each hydrometeor within a cloud system. Microphysics schemes parameterize rates for specific processes such as drop evaporation, but their inherent assumptions lead to uncertainty in model solutions which are often difficult to understand and quantify. The Bayesian Observationally Constrained Statistical-Physical Scheme (BOSS) is a recently-developed bulk rain microphysics scheme with no predefined drop size distribution (DSD) shape and few assumptions made about the process rate formulations. Process rates are written as generalized power functions of a flexible choice in prognostic DSD moments (related to bulk quantities such as mass concentrations). The corresponding parameters are constrained directly with observation using Markov chain Monte Carlo, allowing BOSS to learn microphysical information directly from observations while simultaneously quantifying parametric uncertainty. The process rate formulations in BOSS can be made systematically more complex, which allows us also to track down sources of structural uncertainty.

In this study, we use a detailed bin microphysics scheme as “truth” to generate the constraining observations synthetically, which include profiles of polarimetric radar variables (Z_H, Z_DR, K_DP) and vertical fluxes of prognostic DSD moments at the surface. An error analysis shows that BOSS produces process rate profiles similar to those of a bin scheme. We explore the parametric uncertainty of BOSS through testing how the number of constraining observations affect the posterior parameter probability distributions. We show that increasing the number of constraining simulations leads to better convergence of the parameter distributions for most processes. Increasing the number of prognostic DSD moments from two to three is also shown to improve parameter convergence. Cases where the posterior parameter distributions vary regardless of the number of constraining observations indicates structural error in the model. We explore correlations between the parameters and process rate errors to determine the physical sources of this structural error.